@inproceedings{hazra-etal-2026-safetutors,
title = "{S}afe{T}utors: Benchmarking Pedagogical Safety in {AI} Tutoring Systems",
author = "Hazra, Rima and
Ghuku, Bikram and
Marchenko, Ilona and
Tokarieva, Yaroslava and
Layek, Sayan and
Banerjee, Somnath and
Stoyanovich, Julia and
Pechenizkiy, Mykola",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1917/",
pages = "40864--40896",
ISBN = "979-8-89176-462-0",
abstract = "Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation - failing to capture whether a model is simultaneously pedagogically effective and safe across sustained interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn{'}t reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7{\%} to 77.8{\%}. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn ``safe/helpful'' results can mask systematic tutor failure over extended interaction."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hazra-etal-2026-safetutors">
<titleInfo>
<title>SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rima</namePart>
<namePart type="family">Hazra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bikram</namePart>
<namePart type="family">Ghuku</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ilona</namePart>
<namePart type="family">Marchenko</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yaroslava</namePart>
<namePart type="family">Tokarieva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sayan</namePart>
<namePart type="family">Layek</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Somnath</namePart>
<namePart type="family">Banerjee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Julia</namePart>
<namePart type="family">Stoyanovich</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mykola</namePart>
<namePart type="family">Pechenizkiy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation - failing to capture whether a model is simultaneously pedagogically effective and safe across sustained interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn’t reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn “safe/helpful” results can mask systematic tutor failure over extended interaction.</abstract>
<identifier type="citekey">hazra-etal-2026-safetutors</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1917/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>40864</start>
<end>40896</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems
%A Hazra, Rima
%A Ghuku, Bikram
%A Marchenko, Ilona
%A Tokarieva, Yaroslava
%A Layek, Sayan
%A Banerjee, Somnath
%A Stoyanovich, Julia
%A Pechenizkiy, Mykola
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F hazra-etal-2026-safetutors
%X Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation - failing to capture whether a model is simultaneously pedagogically effective and safe across sustained interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn’t reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn “safe/helpful” results can mask systematic tutor failure over extended interaction.
%U https://aclanthology.org/2026.findings-emnlp.1917/
%P 40864-40896
Markdown (Informal)
[SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems](https://aclanthology.org/2026.findings-emnlp.1917/) (Hazra et al., Findings 2026)
ACL
- Rima Hazra, Bikram Ghuku, Ilona Marchenko, Yaroslava Tokarieva, Sayan Layek, Somnath Banerjee, Julia Stoyanovich, and Mykola Pechenizkiy. 2026. SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 40864–40896, Budapest, Hungary. Association for Computational Linguistics.